Blog

Oxford researchers found that while AI models like GPT-4 correctly diagnose medical conditions 94.9% of the time in isolation, real people using the same AI achieve less than 34.5% accuracy—no better than Google. The failure point isn't the AI's knowledge but the conversation itself, revealing a critical flaw in how medical chatbots are currently deployed.

Researchers discovered that the H3K27M mutation in deadly childhood brain tumors drives excessive lactate production, which directly activates DNA synthesis machinery through lactylation of the enzyme NME1. This finding enabled development of a deuterium-based MRI technique that can visualize tumor metabolism at clinical field strength and detect treatment response early—potentially transforming how we monitor these aggressive cancers.

Researchers tracked 34 cutaneous T-cell lymphoma patients using comprehensive multi-omics and serial sampling, creating the first single-cell resolution map of how cancer evolves to escape treatment. The study identified specific recurrent mutations — including STAT3 D661Y driving HDAC inhibitor resistance and EZH2 alterations targetable with existing drugs — providing a roadmap for genome-guided therapeutic decisions and improved disease monitoring.